2020

Humberto Saenz, University of Texas, Rio Grande Valley
Hongkai Yu, Cleveland State University
Lingtao Wu, Texas A&M Transportation Institute
Xuesong (Simon) Zhou, Arizona State University

With the wide use of smart phones, distraction has become a major safety concern to roadway users. The distracted phone-use behaviors among pedestrians, like Texting, Game Playing and Phone Calls, have caused increasing fatalities and serious injuries. With the increasing usages of driver monitor systems on intelligent vehicles, the distracted driver behaviors can be efficiently detected and warned. However, the research of phone-related distracted behaviors by pedestrians has not been systemically studied. It is desired to improve both the driving and pedestrian safety by automatically discovering the phone-related pedestrian distracted behaviors. In this paper, we propose a new computer vision-based method to detect the phone-related pedestrian distracted behaviors from a view of intelligent and autonomous driving. Specifically, we design the first end-to-end deep learning based Two-Branch Convolutional Neural Network (CNN) for this task. Taking one synchronized image pair by two front on-car GoPro cameras as the inputs, the proposed two-branch CNN will extract features for each camera, fuse the extracted features and perform a robust classification. This method can also be easily extended to video-based classification by confidence accumulation and voting. We propose a new benchmark dataset of 448 synchronized video pairs of 53,760 images collected on an intelligent vehicle in Texas for this research problem, where the proposed method achieved an overall best classification accuracy of 84.3% on the new benchmark when compared to other comparison methods.

Christos Katrakazas, Technische Universitat Munchen
Vishal Mahajan, Technische Universitat Munchen
Constantinos Antoniou, Technische Universitat Munchen

Highway safety has approached significant research interest in recent years, especially as innovative technologies such as Connected and Autonomous Vehicles (CAVs) are fast becoming a reality. Identification and prediction of driving intention is fundamental for avoiding collisions as it can provide useful information to drivers and vehicles in their vicinity, in order for collisions to be avoided. However, the state-of-the-art in manoeuvre prediction requires the utilization of large labelled datasets, which demand significant amount of processing and might hinder real-time applications. In this paper an end-to-end machine learning model for predicting lane change manoeuvres from unlabeled data using a limited number of features is developed and presented. The model is built upon a novel comprehensive dataset (i.e. highD) obtained from German Highways with camera-equipped drones. Density-based clustering is used to identify lane changing and lane keeping manoeuvres and a Support Vector Machine (SVM) model is then trained to learn the boundaries of the clustered labels and automatically label new raw data. The labelled data are then becoming input to a Long Short-Term Memory (LSTM) model which is used to predict manoeuvre class. The classification results show that lane changes can efficiently be predicted in real-time with an average detection time of at least 3 seconds with a small percentage of false alarms. The utilization of unlabeled data and vehicle characteristics as features increases the prospects of transferability of the approach and its practical application for highway safety from researchers and practitioners.

Anik Das, University of Wyoming
MD Nasim Khan, University of Wyoming
Mohamed Ahmed, University of Wyoming

Lane change has been recognized as a challenging driving maneuver and a significant component of traffic safety research. Developing a real-time continuous lane change detection system can assist drivers to perform and deal with complex driving tasks or provide assistance when it is needed the most. This study proposed trajectory-level lane change detection models based on features from vehicle kinematics, machine vision, roadway characteristics, and driver demographics. To develop the models, the SHRP2 Naturalistic Driving Study (NDS) and Roadway Information Database (RID) datasets were utilized. The lane change detection models were trained, validated, and comparatively evaluated using three Machine Learning algorithms including Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN). The results revealed that the highest overall detection accuracy was found to be 88.9% using the RF model when all the features were included in the model. Similarly, it was observed that the RF model had the highest overall detection accuracy of 79.9% while excluding machine vision-related features. Moreover, the highest overall detection accuracy of 77.1% using ANN model was observed considering only vehicle kinematics-based features, indicating that the proposed model could be utilized with reasonable accuracy when other data are not available. The findings of this study may provide appropriate guidance to transportation researchers in conducting lane change-related research. The study could also provide valuable insights regarding drivers’ aggressiveness in terms of lane change behavior in a Connected Vehicle (CV) environment.

Cheuk Ki Man, Loughborough University
Mohammed Quddus, Loughborough University
Athanasios Theofilatos, Loughborough University
Rongjie Yu, Tongji University
Maria-Ioanna Imprialou, Atkins

In recent years, real-time crash prediction has been heavily studied amongst researchers. Traditional real-time crash prediction models adopted matched case-control sampling (i.e. balanced data) for crash prediction. Various statistical and machine learning methods have been employed in order to improve predictability of these models. However, crash prediction results from matched case-control sampling have been criticised as being biased and erroneous. In this paper, full imbalanced dataset is utilised to reflect the reality for crash prediction. This is challenging as the ratio of ‘crash cases’ against ‘non-crash cases’ is usually high as crash is a rare and stochastic phenomenon. As a result, non-crash cases would dominate the traffic dataset. Therefore, class balancing is needed to enable crashes to be learnt properly. In this study, a state-of-art class balancing method, Generative Adversarial Network (GAN) is adopted. Crash data and traffic data were collected from M1 Motorway in United Kingdom. Crash prediction is then carried out to examine and compare the effect with different amounts of additional synthetic crashes generated by Wasserstein Generative Adversarial Network (WGAN), Synthetic Minority Over-sampling Technique (SMOTE) and Adaptive Synthetic (ADASYN). In this study, all oversampling methods displayed excellent crash prediction results, where SMOTE is the best performing oversampling method. WGAN also displayed a satisfactory result with the best sensitivity of 0.87 at 0.5% false alarm rate. Findings of this study reveal that WGAN is a viable oversampling method to handle extremely imbalanced dataset (with a ratio of over 1:11,000) and oversampling would vastly enhance crash prediction ability.

Meng Li, Southeast University
Zhibin Li, Southeast University
Chengcheng Xu, Southeast University
Tong Liu, Southeast University

The primary objective of this study is to propose a deep reinforcement learning-based driving strategy for individual vehicles to mitigate oscillations and optimize traffic safety in stop-and-go waves. A Deep Deterministic Policy Gradient (DDPG)-based driving strategy, which requires information that are directly obtained by in-vehicle sensors, was proposed for system performance optimization. Two typical scenarios were simulated based on a simulation software (SUMO): i) the leading vehicle slowed down according to a real trajectory data to produce one oscillation; ii) the leading vehicle conducted several abrupt decelerations with various degrees of disturbance to produce multiple oscillations. The DDPG agents interacted with SUMO platform to determine the optimal acceleration of vehicles that can reduce crash risks in various stop-and-go waves. The results showed that the proposed DDPG-based driving strategy successfully reduced the crash risks by 68.9% to 78.4%. The scenarios with different penetration rate of the DDPG agents and in various flow rates were compared to test the effect of the proposed strategy. The DDPG-based driving strategy reduced more crash risks with the increase of penetration rate and this strategy performs better when applied in the scenario with a high flow rate. Lastly, we compared our proposed strategy with those of ACC strategy and jam-absorbing driving (JAD) strategy. Results showed our strategy outperformed other oscillation mitigating strategies in reducing crash risks.